Introduction: the cognon concept

Experiments have shown that cognitive activity can be seen as a transient sequential switching across different metastable states, e.g. see the reviews (Tozzi et al. 2017; He 2018; Michel and Koenig 2018; Rabinovich et al. 2020). These states arise on three levels of the brain hierarchy: neuronal, micro-network, and large-scale functional networks. Cognitive dynamics involve many different neural processes and resources including perception, memory, decision making, attention and emotion. In spite of the enormous variety of these processes, which involve multiple brain regions and coordination mechanisms (Tognoli et al. 2021), they demonstrate an amazing universality from the dynamical point of view (Rabinovich et al. 2012b): Within a wide range of measurement modalities and spatiotemporal scales, one observes sequential dynamics between transient metastable states.

Cognitive dynamics can therefore be interpreted as robust sequential switching described by associated metastable sets on different levels of their winnerless competition spatio-temporal activity (see Box 1). The dynamics of these processes can also be described by basic universal models from this viewpoint. In this paper, we introduce the concept of cognon as the unit of thought, which is represented by a robust finite chain of metastable states. Many mathematical models naturally provide the existence of not only metastable states, but also the associated hierarchy of their chains, which serve to describe cognon dynamics. Below we will illustrate this with generalized Lotka–Volterra and Ginsburg–Landau models.

Neural dynamics of memory and binding

Episodic memory (EM)—as an example of a cognitive process—contains various details of an event, such as the objects or people involved (“what”), the spatial setting (“where”) and the temporal sequence (“when”) in which the event unfolded. To create an EM representation, the spatiotemporal information must be remembered as a coherently bounded sequence of episodes. For example, the art of music involves balance between surprise and predictability. It is possible to consider music generation as a three dynamic modality process: melody, harmony and rhythm.

Episodic memory is consciously recollected in spatio-temporal neural activity patterns related to personally experienced groups of events, i.e., episodes (e.g. see Bergström et al. 2013; Ekstrom and Ranganath 2018). Episodic memory retrieval is a dynamic process that draws upon the sequential ability to reconstruct past experiences from corresponding cues. The neural substrates of these abilities are engrams, which are sets of basic units of memory in the form of mini-networks of neuronal clusters (Kitamura et al. 2017).

Binding is a key dynamical mechanism for the implementation of autobiographic episodic memory (Gilbert et al. 2014; McGatlin et al. 2019); see also Box 2. Binding is the process by which frequently repeated segments of temporal inputs are concatenated into single conceptual units that are easy to process (Gobet et al. 2001). Such processing of information is fundamental to time-series analysis in biological and artificial neural computation systems. The brain efficiently acquires integrated information from various modality streams in an unsupervised manner.

Cognon and brain rhythms

Cognitive brain activities can be represented by the dynamics of two qualitatively different components: (1) the specific activity of cognitive networks, and (2) the overall dynamics of continuous oscillatory fields, i.e., brain rhythms. These components process cognitive information in different ways and are often considered independently. However, their mutual interaction through synchronization and desynchronization creates a universal processor with a unique ability to operate with different forms of cognitive mechanisms (see Box 4). The study of such interactions can also help to the assessment of mental disorders (Rabinovich and Varona 2017; Fingelkurts and Fingelkurts 2019). Their associated metrics and parameters that control their relationship can be convenient biomarkers and contribute to novel rehabilitation protocols (Latorre et al. 2019).

To our best knowledge, a consistent mathematical model based on the mutual interaction of both these components does not exist yet. We suggest here for the first time a model of “cognon-field” information dynamics, which we describe by linking a formulation of heteroclinic dynamics—saddle invariant sets in the cognitive phase space connected by heteroclinic orbits—for competitive cognitive interactions and complex Ginzburg–Landau fields (see Box 4).

To build a nonlinear theory of, for example, autobiographic memory dynamics, it is necessary to create a universal scale-free model—a canonical model—of cognitive dynamical processes. We suppose that the canonical model has to satisfy the following conditions: (a) The equations have to be written for variables that can represent the evolution of brain oscillatory clusters in their temporal coherence and have to have solutions that correspond to metastable patterns in the brain; (b) the model is based on winnerless competitive dynamics—a nonlinear process of interaction of many agents that guarantees the sequential switching between metastable states and the robustness of transients, (c) the model is an open dissipative system where inhibition is balanced by excitation, (d) the model’s dynamics have to be sensitive to the incoming information, and (e) be able to describe closed heteroclinic chain dynamics. The mathematical image of the cognon is a Stable Heteroclinic channel (SHC) that consists of a sequence of metastable states (see Box 1).

The reduction of high-dimensional brain data to a low-dimensional cognitive space can be motivated by empirical observation. There are a number of experiments that have illustrated the low-dimensionality of cognitive dynamics when it is governed by sensory stimuli (Shine et al. 2019). Formally, this means that large amounts of data can be represented by the dynamics of a reduced number of spatiotemporal patterns—or modes—e.g., using spatiotemporal decomposition techniques, see Banerjee et al. (2012), Pinotsis et al. (2014) and Glomb et al. (2017).

In terms of EM, the formation and retrieval of event memories are implemented by collaborative dynamics between the neocortex and the hippocampus, as can be observed by the analysis of neocortical alpha/beta frequency desynchronization and hippocampal theta/gamma frequency synchronization (Griffiths et al. 2019). In this task, the neocortex processes event-related information and the hippocampus binds this information. Such brain rhythm analyses indicate that a bidirectional information exchange between the neocortex and the hippocampus is fundamental for the formation and retrieval of episodic memories.

The learning dynamics—forcing of sequential events by the environment—activates the chain of engrams in time. On the retrieval stage, this chain of engrams replays robustly through the sequential competitions between individual engrams in time (Rashid et al. 2016; Rao-Ruiz et al. 2019; Takamiya et al. 2020) and thereby reconstructs the original sequence of events.

As an example, let us consider creative cognition or goal directed self-generated creative thought (Beaty et al. 2016). First, it necessary to consider the dynamic interaction of large-scale brain networks and their constituting processes that participate in creative tasks. These dynamical processes are: (1) creativity, idea generation, and elaboration, (2) sequential working memory, (3) attention, and cognitive correlation and control. Creative idea generation is based on interactions including frontal-central as well as frontal–temporal networks (Rominger et al. 2020). Variables, i.e., amplitudes and phases that describe the different modes of this network form the cognitive phase space where there exists a heteroclinic structure based on metastable states (Rabinovich et al. 2008b), the mathematical image of cognons. The complexity of the functional connection matrix (see Eq. (2) in Box 4)—depends on personality and the stage of the creativity process (Fink and Benedek 2014). At the same time, the creativity dynamics, including its constituting functional components, like working memory (de Vries et al. 2020), are controlled by brain rhythms in the alpha frequency range. Alpha power helps estimating the creativity level of ideas, and is responsible for the functional correlation of large scale brain activity (Benedek and Fink 2019). Alpha oscillations also play an important role in the organization of divergent thinking and the performance on the alternative uses task (Agnoli et al. 2020), and serial order effect (Kraus et al. 2019). In our model, the power of brain oscillations can control the elements of the connection topology (matrix ςij in (2)), which determine the existence of cognons. At the same time, the cognon influences the generation of alpha oscillations making the basic model with such feedback self-consistent.

Conclusions

The concept of a cognon captures the sequential nature of essential cognitive processes, which can also be seen within the framework of a generative model linking continuous and discrete time descriptions of neural activity and its associated behavior (Rabinovich and Varona 2018; Parr et al. 2023). The hierarchical multilevel timing architecture of basic cognon models is a convenient way for analyzing sequential binding phenomena of cognitive dynamics. It can be a sequence of events in the episode or a group of symbols (words) forming a thought. In contrast, with the chunking process of one modality, the process of binding of different perception features—or modalities—requires temporal coordination of parallel transient modalities (Fingelkurts et al. 2003b). For such description, the number of different layers in the model has to coincide with the number of modalities.

The multilayer dynamics of (4) and (5) is also adequate to model decision-making (DM) processes in the case when DM is viewed as a choice between different binding modalities modulated by the environment. This binding occurs in many decision-making tasks related to behavior.

Different physiological signals, as recorded by in vitro, in vivo electrophysiological experiments, fMRI and EEG, capture the same brain activity from different time and spatial scales, and their dynamical interaction (Van De Ville et al. 2010; Bassett and Sporns 2017; Avena-Koenigsberger et al. 2017). Interlaced robust sequential informational processes are observed in subcellular, cellular, small network and large network interactions, including systems interaction. A basic dynamical model has to describe the simultaneous temporal evolution of these signals and their mutual interaction depending on the specific cognitive goal and environment condition. Building sequences of sequences with such dynamical objects naturally gives rise to coordinated hierarchical neural phenomena at multiple description levels, from microscopic to macroscopic information flows, from sensory encoding to cognitive decision making.

The cognon approach described by the above discussed model consists of two groups of equations, i.e., equations for continuous spatio-temporal fields, and equations of cognon dynamics. They interact by mutual modulation of the control parameters including mutual excitation and inhibition. Of course, hypotheses about the robustness and reproducibility of solutions of (4) have to be proved in the future. However, available experimental data about the existence and reproducibility of metastable informational patterns in active brains (Tognoli and Kelso 2014; He 2018; Roberts et al. 2019) support this hypothesis and can guide new experiments, including those related to pathological states (Rabinovich and Varona 2017). The concept of cognon, as the basic unit of cognitive information can help to bridge the gap between theoretical formalisms of cognitive dynamics, physiological measurements and information programming of behavior. Additional impact of this concept can include the realm of artificial intelligence and, in particular, quantum inspired artificial agents (Huber-Liebl et al. 2022).